Platform for distributed optimization experiments using OpenMPI

Introduction

This project is a platform for distributed optimization expriments using OpenMPI. This platform is used to implement the expeirments of our recent papers [1][2]. Please cite them if this code helps you. :)

[1] Zhouyuan Huo, Heng Huang, Asynchronous Stochastic Gradient Descent with Variance Reduction for Non-Convex Optimization. AAAI, 2017

[2]Zhouyuan Huo, Bin Gu, Heng Huang, Decoupled Asynchronous Proximal Stochastic Gradient Descent with Variance Reduction. arXiv

Installation

Clone this repository from github.
run ./install

Remark: ./install is only used for first time installation. After you modify the code and want to compile, just run make in build/ directory.

How to use on localhost

Run linear model.

Generate distributed datasets. Go to build directory, and run ./splitdata ../data/covtype_binary 4 0. It will generate 4 data files in data/covtype_binary_split/ directory. ./splitdata will transform libsvm format data to armadillo format data, and distributes it to multiple files. run ./splitdata directly to see how to use it.
Run To run do

mpirun -np 5 ./mpiplatform -logistic_l2_l1 -num_workers=4 -data_file="absolutepathto/data/covtype_binary_split/" -print_loss_per_epoch -d1=54 -learning_rate=1e-1 -n_epochs=100 -mini_batch=100 -in_iters=1000 -svrg -max_delay=10

Run fully connected neural network model.

Generate distributed datasets. Go to build directory, and run ./splitdata ../data/covtype_multiclass 4 0. It will generate 4 data files in data/covtype_multiclass_split/ directory.
Run To run do

mpirun -np 5 ./mpiplatform -fcn -num_workers=4 -data_file="absolutepathto/data/covtype_multiclass_split/" -print_loss_per_epoch -d1=54 -d2=20 -d3=7 -learning_rate=1e-3 -n_epochs=100 -mini_batch=10 -in_iters=1000 -svrg -max_delay=10

How to use on AWS

Open an account on Amazon Web Services (AWS).
Launch instances on AWS EC2. Configure: (1) Choose AMI: Ubuntu Server, (3) Configure Instance: Number of instances 5, (6) Configure Security Group: Type: All TCP. Then click launch and download a key named FirstKey.pem.
Setup the public key. mv Firstkey.pem ~/.ssh/ ＆ chmod 400 ~/.ssh/Firstkey.pem
Generate compressed data. Go to data/ directory, and run ./compress_data.sh absolutepashto/data/covtype_binary_split/ 4.
Generate compressed code. Download our project from github again, copy Firstkey.pem to the repository and compress it. run tar cvf aws_code.tar MPIPlatform
Upload and install. Open upload_install_split.sh and fill in the hosts part with public ip of 5 machines we just launched.

hosts="
54.175.225.66
54.174.70.95
54.211.115.252
52.91.166.132
54.175.107.232
"

First machine works as server. Then upload and install by running

./upload_install_split.sh absolutepathto/aws_code.tar  absolutepathto/covtype_binary_split/

7 . Login to server and setup. To login the server, do

ssh -i ~/.ssh/Firstkey.pem ubuntu@54.175.225.66

Go to bulid/ and create a hostfile like

54.175.225.66
54.174.70.95
54.211.115.252
52.91.166.132
54.175.107.232

8 . Run the program, do

mpirun -hostfile hostfile ./mpiplatform -logistic_l2_l1 -num_workers=4 -data_file="/home/ubuntu/" -distribute -print_loss_per_epoch -d1=54 -learning_rate=1e-1 -n_epochs=100 -mini_batch=100 -in_iters=1000 -svrg -max_delay=10

Disclaimer

This repository uses code from Cyclades, we borrow the framework from this project. And we use code from LibSVM to read libsvm data and transform it to distributed armadillo format.

slowbull/MPIPlatform